A Hybrid Modelling of a Water and Air Injector in a Subsonic Icing Wind Tunnel
††thanks: This work was supported by the Dispositif Recherche of Métropole Rouen Normandie through the COPOGIRT project.
For this reason and the purpose of Open Access, the aut (2024)
César Hernández-Hernández, Thomas Chevet, Rihab el Houda Thabet, and Nicolas LangloisThe authors are with Université de Rouen Normandie, ESIGELEC, IRSEEM, 76000 Rouen, France{cesar.hernandez, thomas.chevet, rihab.hajrielhouda, nicolas.langlois}@esigelec.fr
Abstract
The study of droplet generation in wind tunnels in conducting icing experiments is of great importance in determining ice formation on structures or surfaces, where parameters such as Liquid Water Content (LWC) and Median Volumetric Diameter (MVD) play a relevant role.The measurement of these parameters requires specialised instrumentation.In this paper, several experiments have been carried out in a subsonic wind tunnel facility to study the parameters that are part of the icing process in structures.Furthermore, a mathematical modelling of the constituent subsystems of the plant study that allow us to have a comprehensive understanding of the behaviour of the system is developed using techniques based on first principles and machine learning techniques such as regression trees and neural networks.The simulation results show that the implementation of the model manages to obtain prominent expected values of LWC and MVD within the range of values obtained in the real experimental data.
Index Terms:
Mathematical model, subsonic icing wind tunnel, liquid water content (LWC), median volumetric diameter (MVD), regression tree, neural network.
I Introduction
In the process of structural icing, many parameters are involved, for instance, temperature, wind velocity, air pressure, humidity, liquid water content (LWC) and median volumetric diameter of water droplets (MVD).LWC is normally expressed as the number of grammes of liquid water per cubic meter of air.It represents the amount of supercooled water droplets that can impact the aircraft surface in a given air mass.The diameter of the water droplets is usually characterised as the median volumetric diameter normally expressed in micrometre, and LWC and MVD are closely related [1].Determining the LWC detection of the distribution of droplets produced by a spray system in a wind tunnel is of great importance to prevent ice formation.MVD is usually measured with different instruments, which generally require interaction with the droplets under sensitive conditions and are prone to fail under very icy conditions.Therefore, LWC and MVD are important factors in structural icing.
The possibility of ice formation under atmospheric conditions has become a major problem in different areas, such as in electrical networks [2], wind turbines [3, 4, 5], aircrafts [6, 7] and helicopters [8].
Considerable work has been done to study the physics and nature of ice formation [9], where LWC and MVD are highly essential parameters related to the occurrence of ice [10].In [11], the authors integrate three measurement devices into a wind tunnel to measure particle size distribution (PSD), MVD, and LWC.Moreover, in [12] a -nearest neighbour-based model of ice intensity has been developed from wind velocity, LWC and MVD measurements from two instruments.In [13], variables related to LWC are studied to determine the effect on engine conditions due to the change in LWC and atmospheric conditions.
In this work, we propose a hybrid nonlinear mathematical model of a subsonic icing wind tunnel and its constituent subsystems.The model is developed and implemented in Matlab/Simulink [14], and obtains values that are within the range of experimental test values, where there are atmospheric conditions measurements that allow LWC values between and and MVD between and .To develop the hybrid model, we use models based on the lumped parameter approach [15] and data-driven models for the variables involved in the process.We show simulation results where different scenarios are considered to study the variables of interest LWC and MVD.
The remainder of the paper is organised as follows. Section II presents the information about the experimental plant.Section III presents the mathematical modelling of the plant.Section IV presents results obtained with the plant modelling implemented in simulation.Finally, Section V presents conclusions and future work.
II Experimental setup
In this work, we study an experimental plant consisting in a closed-loop subsonic wind tunnel associated with a water and air injection system.Before describing the modelling process, we introduce, in this section, information about this plant.To do so, we give a brief presentation of the two constituent parts.
The first part of the facilities is the water and air injection system.This system is composed of a water and an air tank, each connected to 12 conduits.Each of the 24 conduits possesses its control valve to regulate the flow.Finally, each water conduit is paired with an air conduit into a nozzle that injects a mixture of water and air into the second part of the facilities.
The second part is the wind tunnel itself.This tunnel is associated with a cooling chamber in order to maintain a negative temperature during experiments.These negative temperatures coupled with the injection system described before aim to generate an ice fog inside the wind tunnel to test defrost equipment.Control panels, allowing to adjust various parameters in the plant, are connected to it.
A schematic view of this plant is given in Figure 1.
II-AEquipment description
After this brief introduction, we now provide more details on the equipment of the icing wind tunnel facility.
The water tank is considered to be a closed cylindrical tank high with a radius of approximately .It has a capacity of .This tank is pressurised by constant air injection at a pressure of , and is heated by five resistors (four of and one of ).In addition, a pressure regulator valve is installed for safety purposes.In terms of sensors, the tank is equipped with a type J thermocouple, which is capable of measuring temperatures between and .Finally, a sensor is present to measure the water level inside the tank.For its part, air tank is assumed to work as a blower, that is, the air coming from a pressurised air line is heated at a constant chosen temperature and injected into the system at a pressure of .
As mentioned above, each tank is connected to 12 conduits in which the flows are controlled by solenoid proportional valves.The 12 air valves are of type ASCO SCG202A001V and the 12 water valves are of type ASCO SCG202A051V.These valves are controlled by proportional integral (PI) controllers.Furthermore, each bus is equipped with flowmeters that can be used to measure air and water flows, as shown in Figure 2(a).
The droplets are accelerated and injected into the test section by nozzles of type SUJ16 pneumatic atomizers.Each nozzle has a heating patch that is used to maintain the temperature above to prevent the water droplets from freezing before injection in the test section.
Figure 2(b) shows the front part of the test section.In this area, the velocity and temperature of the wind are controlled throughout the experiment by external control systems.
As detailed previously, the entire plant is used to test defrost systems.Therefore, it is necessary to control as accurately as possible the LWC and MVD of water droplets injected to perfectly characterize the tested devices.For the present works, this means that we require a precise model giving these LWC and MVD as a function of the injection system’s parameters.To do so, in previous experiments, a JRT measuring instrument was placed in front of the test section (see Figure 2(c)) to accurately measure LWC and MVD.The next section describes the data set obtained in these experiments.
II-BExperimentation and description of the data collected
To obtain LWC and MVD data for modelling, thirty experiments were carried out.Data were collected from sensors installed throughout the plant as described in the previous section.The data collected during these experiments are
•
the temperatures in the test section , the water tank , the air tank , and the nozzles , all in Celsius degrees;
•
the wind velocity in the test section in ;
•
the LWC in the test section in ;
•
the MVD in the test section in ;
•
the volumetric air flow per bus in and volumetric water flow per bus in .
The liquid droplets produced by the atomisers have a size of to and are injected at a speed of to .In the test section, these droplets are supercooled at temperatures from to .In such experiments, we obtain LWC values ranging from to , while MVD values range between and .Table I shows the main statistical indices of these experiments, calculated over thirty values for each variable, and Table II shows the correlation between the variables involved.
mean
standard deviation
minimum
maximum
The data collected in these former experiments is then used in the following section to get a numerical model usable for control and simulation.
III Modelling methodology
The development of mathematical models in icing wind tunnels is highly complex due to the large number of variables and complex processes involved.In this section, a hybrid nonlinear model is developed based on lumped parameters approach and machine learning techniques.With this strategy, we provide a mathematical representation for the water and air tanks, the flow control valves, the nozzles, and the test section, aiming to link the LWC and MVD with these components’ parameters.
III-AWater tank
As presented in Section II-A, one of the first elements of the plant we study in this paper is a water tank.To study the dynamic behaviour of this water tank, illustrated in Figure 3, we split the modelling into two parts.On the one hand, mass balance is used to model the level of water inside the tank.On the other hand, energy balance is used to model the influence of heat sources on the temperature of the water.
Water level
The water tank has a cross-sectional area , in , and a height , in .As mentioned in the previous section, the tank is pressurised, at all times, with a constant pressure .At the bottom of the tank is an opening having a cross-sectional area , in .When water is injected into the test section, this opening allows the tank to empty into the opened water buses.At this moment, the surface of the water in the tank flows down at a velocity , while it flows at a velocity through the orifice, both velocities being in .Finally, when emptying, the height of the liquid will evolve from an initial value , in , to adopt a value , in .
where is the water mass inside the tank, and and are the input and output fluids mass flow rate inside the tank.
Remark 1.
In the following, the input mass flow rate is a control variable that is chosen by the user depending on the operational needs.In addition, is the water mass flow rate injected into the test section.
In the tank, the overall mass can be divided into a mass of water and a mass of air , so that
(2)
However, given the tank’s volume, we have .It follows that the mass of the air can be considered constant, so that
(3)
As described before, the tank is cylindrical in shape and , with the water density in , so that
(4)
Therefore
(5)
where , in , is the water density.
Water temperature
In terms of temperature, the energy change in volume control is due to several elements.The first source of change is the power provided by the heating system , in .The second source are the losses caused by leaks to the outside of system where , in , is a constant characterising the leaks, and , in , is the temperature in the water tank.The third source is the input energy provided by the incoming mass , where , in , is the water’s mass specific heat, and , in , is the temperature of the incoming water.Finally, the last source is the output energy provided by the outgoing mass , where is the temperature in the water tank expressed in Kelvin.Then, considering the energy , in , of liquid water at a certain temperature, we have
(6)
The energy rate of change is then
(7)
and, using (1) into (7),
(8)
Given what we have described above, we have
(9)
so that
(10)
or
(11)
III-BAir tank
The second element of the plant described in Section II-A is the air tank.In this Section, this air tank is described as a blower.It follows that its geometry is not relevant for modelling purposes, and the important elements are the density of the air blown and its temperature.
As mentioned in [16], following the law of conservation of matter [17, 18], the air tank mass variation is equal to the mass flow inlet minus the mass flow injected into the test section .In addition, the mass of air in the tank can be written where , in is the air density, and , in , is the volume of the air tank.It follows that the variation of air density in the air tank is
(12)
With regard to temperature, the same methodology that has been used for the water tank has been followed.Therefore, the evolution of the air temperature in the air tank is
(13)
where , in , is the temperature of air in the air tank, , in , the temperature of the air injected inside the tank, , in , the power provided by the heating system, is the air’s mass specific heat, in , and , in , a constant characterizing the leaks.
III-CFlow control valves
The elements placed right after the water and air tank are flow control valves allowing to control the flow of the mix of air and water injected into the test section of the plant.As mentioned in Section II-A, these water and air control valves are proportional solenoid valves.Table III shows the main characteristics of these valves.
Orifice size
Flow factor ()
Operation pressure ()
Water valve
According to the technical report [19] of the water valve provided by the manufacturer, the volumetric flow of water through a valve is a function of the valve’s opening area , in , the discharge coefficient , the pressure difference , in , in the valve, and the density of water , in , and is given by
(14)
The discharge coefficient is obtained as
(15)
where is the flow factor, in , given by the manufacturer, and , in , is the diameter of the valve’s orifice.The difference of pressure is obtained as
(16)
where is the pressure drop coefficient and , in , is the velocity at which water flows through the valve.The pressure drop coefficient is given by
(17)
Air valve
Through the air valve, the volumetric flow is given by [20, 21]
(18)
where is the ratio of specific heats, is the air-gas constant, , in , is the air pressure in the air tank pressure, and , in , is the air pressure outside the control flow valve.
III-DNozzles
Downstream the flow control valves we just described are nozzles at which level the air and water coming from the tanks are mixed and injected into the test section.Nozzles are elements that are used to increase wind velocity based on their geometry.As shown in Figure 1, twelve nozzles are installed at the entrance of the test section.They are arranged in three columns of four nozzles each.This Section aims to study the air velocity that each nozzle provides to the test section, as well as the temperature in the nozzles.A temperature higher than zero must be maintained so that there is no freezing in this part and the experiments can occur safely.As can be seen in Table I, temperatures between and are maintained in this area.
Velocity at the nozzles’ output
With regard to velocity, a very simplified model using Bernoulli’s continuity equation is used to determine the velocity provided by the nozzles towards the test section.Bernoulli’s equation can be viewed as a conservation of energy law for a flowing fluid.Figure 4 presents a schematic view the nozzles’ geometry.The input surface is taken times bigger than the outlet surface .Hence, the velocity at the outlet of the nozzles is given by
(19)
where and , in , are respectively the inlet and outlet surfaces, and and , in , are the velocity of the air coming from the air tank and of the mix of air and water injected in the test section.It follows that
(20)
Remark 2.
As usual for this type of equipment [11], the velocity of the water injected inside the nozzle is not taken into account, as it is negligible.It is the pressure provoked by the air velocity that leads to the spraying of a mix of air and water at the output of the nozzle.
Data-driven model of the mix’s temperature in the nozzles
Temperatures in the nozzles are not measured by a sensor.However, it can be estimated by obtaining a mathematical model that depends on other known variables.To do so, we performed a statistical data analysis to identify variables that have a greater impact on the temperature of the mixture of air and water in the nozzles.We then extract a polynomial model of depending on the identified variable.
To determine which of the system’s variables can be used to express , we use mutual information gain (MIG) [22] and f-score [23].Figure 5(a) presents the mutual information gain of the system’s variables with .It shows that the temperature of the nozzles depends on other temperature values, where the most important variables are , , and , followed by , , , and that have a minor influence on the temperature of the mixture in the nozzle .In addition, according to the f-score, presented in Figure 5(b), and have a greater influence than the variables , , and .Therefore, to ensure that the nozzle temperature only depends on other temperature values, we chose not to use and to estimate .As a result, the variables , , and have been used to obtain different models based on the data for the evolution of .
From this analysis, a regression tree, a neural network [24], and a polynomial model [25] for are computed and compared with each other.The mix’s temperature in the nozzles is given as a function of only , , and .As described in Section II-B, a low number of experiments ( in total) were performed, leading to a low number of data.It follows that all acquired data has to be used to estimate the models.In order to avoid overfitting, this is done using a cross-validation of .The models are obtained by minimising the mean square error (MSE), [26].Finally, the models are validated on the experiments themselves.
A simple form is imposed to determine the polynomial model giving as a function of , , and .Therefore, minimizing the MSE, we obtain
(21)
where all the temperatures are expressed in degree Celsius.The regression tree giving the value of is presented in Figure 6.It has to be read as tests on the variables , , and , expressed in degree Celsius.Depending on the position of these variables with given thresholds, indicated on the tree’s branches, we obtain a value, in degree Celsius, for .
For the neural model of , we optimize the properties of the neural network among the following possibilities:
•
the number of hidden layers is either , , or ;
•
each hidden layer can contain between and neurons;
•
the activation function is either , , or .
As for the other models, a 5-fold cross-validation is used and the characteristics of the neural network are obtained through minimisation of the MSE.The optimal characteristics of the neural model of are presented in Table IV.
Hidden layers
2
Neurons per layer
10 & 4
Activation function
sigmoid
The proposed models are first validated on the data from previous experiments described in Section II-B.To do so, Figure 7(a) compares the values of obtained with the polynomial model (in red), the regression tree (in blue), and the neural model (in green) with the values of from the 30 experiments in the data set.We see that the regression tree and the neural model behave better than the polynomial model.Figure 7(b) confirms this behaviour as it shows the error between each model and the data from the experiments, the colours being the same as in Figure 7(a).To further validate the models, we present the overall mean square error and mean absolute error (MAE) [26] for the different models in Table V.It confirms that the regression tree and the neural model perform better than the polynomial model, with the neural model performing slightly better than the tree.However, due to the low quantity of data available in the data set, the regression tree shown in Figure 6 has a low number of possible states for .Therefore, we decide not to use it in the following as it might not work well in a more general simulation process.
Model
MSE
MAE
Polynomial
66.70
6.20
Regression tree
10.68
2.03
Neural network
9.57
1.97
III-ETest section
As already exposed, the goal of the plant is to inject a mix of air and water into the test section.As explained in Section II-A, the variables of importance are the liquid water content and the median volumetric diameter .Two other variables, the temperature and the wind velocity in the test section are also important.However, they are regulated by external systems, so we consider them as input variables and focus on LWC and MVD modelling.
Liquid water content
According to [11], LWC represents the mixing of the available mass of water within a defined air volume in the test section, given by
(22)
The volume of air flowing through the test section can be calculated by where is the cross-sectional area of the test section.The available mass of water is .It follows that
(23)
The MIG and f-score for the LWC, showing the influence of the plant’s variables over the LWC, are presented in Figure 8.They show that the flow of water plays indeed an important role in the expression of as is implied by the relation (23).However, the influence of the air velocity in the test section is less important according to both the MIG and f-score.Nevertheless, the expression (23) is grounded into other experiments from other testbeds [11].Therefore, we decide to model the LWC with the expression given in equation (23).
Mean volume diameter
As highlighted in the Introduction, the estimation of the MVD is a much more complicated task.In the literature, various measurement devices [11] or machine learning techniques [12] have been used, as mentioned previously.Therefore, from the data described in Section II-B, we follow the same methodology to model the MVD as we did for the temperature in the nozzle as exposed in Section III-D.
We start by obtaining the mutual information gain and f-score to know which variables have the greatest influence on the MVD’s value.The MIG and f-score for MVD are shown in Figure 9.According to the MIG values in Figure 9(a), the variables , , , and are the variables that influence the MVD value.In addition, the variable has a more important influence compared to the others.This observation is in accordance with the f-score presented in Figure 9(b).
Figures 8 and 9 show valuable information, as the LWC does not appear to have a great influence on the value of MVD and vice versa.However, because of the definition of these two variables, if the liquid water content in the droplets is zero, we cannot get a median volumetric diameter.In addition, we already saw that LWC values depend on the water flow , and MVD values depend on the air flow .
It follows, by symmetry, and the influence that it has according to Figure 9, that we decide to express the MVD as a function of .From the previous discussion, we also decide to include the LWC in the MVD’s expression.Finally, although they do not have a great influence on the MVD according to the values in Figure 9, we include the variables describing the conditions in the test section, i.e., the temperature and the wind velocity , as the MVD is measured in the test section.
As for the temperature of the mix of air and water in the nozzles , we want to obtain a polynomial, a regression tree, and a neural model for the MVD.The process for obtaining them is the same as described in Section III-D.
To obtain the MVD polynomial, more complex forms than for are imposed.First, as has the highest influence on the MVD according to Figure 9, we propose a polynomial model where is a factor of all terms composed of all possible combinations of 1 to 3 variables among , , and to the power .Minimizing the MSE with this form, we obtain
(24)
The second form we consider involves all possible combinations of products of 1 to 4 variables among , , , and to the power .Therefore, minimizing the MSE, we obtain
(25)
where the temperature is expressed in degree Celsius.The regression tree giving the value of is presented in Figure 10.It is interpreted in the same way that the regression tree for is interpreted as described in Section III-D.
The neural model of is obtained under the same conditions as the neural model of .Its optimal characteristics are presented in Table VI.
Hidden layers
2
Neurons per layer
2 & 7
Activation function
ReLu
As for the temperature , the proposed models are first validated on the data from previous experiments described in Section II-B.To do so, Figure 11(a) compares the MVD values obtained with the polynomial model from (24) (in brown), the polynomial model from (25) (in red), the regression tree (in blue), and the neural model (in green) with the values of from the 30 experiments in the data set.From the curve, it is difficult to properly separate the different models as they have quite a similar behaviour.Figure 11(b) confirms that as it shows the error between each model and the data from the experiments, the colours being the same as in Figure 11(a).To further validate the models, we present the overall mean square error and mean absolute error for the different models in Table VII.It confirms that the models behave really closely.However, with the values of Table VII, we decide not to use the polynomial model given by equation (24) as it clearly has poorer performance compared to the other models.As for the temperature in the nozzle , we also decide not to use the regression tree, we also decide not to consider the regression tree model given in Figure 10 as, due to the low number of data in the data set, gives a low number of possible states.
Models
MSE
MAE
Polynomial (24)
Polynomial (25)
Regression tree
Neural model
IV Results
In this section, we present results obtained in simulation and discuss the analysis and behaviour of the system.As mentioned previously, it is of great importance to have atmospheric conditions that encourage the study of variables such as LWC and MVD that are related to ice formation.The models developed in Section III are incorporated in interconnected modules in a Simulink model shown in Figure 12, which formed a hybrid model of the plant.A dashboard with different functionalities has been designed to simulate scenarios in which various modes of plant operation can be used.The results shown in this section are from a simulation of with a sampling time of .The conditions considered in the simulation are described below.
•
The variables involved in the simulation process have been executed within the ranges of real conditions as shown in Table I.
•
Different valves have been activated and deactivated throughout the experiment.
•
In the simulation, changes in the temperature of the test section, as well as in the velocity of the test section, have been considered.
•
The temperature () and the velocity () in the test section are considered variables controlled by an external system.
•
The flow through each valve is controlled by a PI controller (in the internal control loop) and the PI acts on the opening and closing area of the valves.
To observe the time response of LWC, a water flow of from initial time to , from second to , and from to , respectively, per conduit (see Figure 13(a)) has been applied.LWC is directly proportional to and inversely proportional to according to the equation (23).Therefore, the changes in LWC shown in Figure 13(c) are due to the amount of water flow, the number of conduits used (Figure 13(e)) and the changes in velocity in the test section (Figure 13(f)).
MVD varies according to the equation (25) where has the strongest influence.As is done to LWC, an air flow per conduit of from the initial time to , from to , and from to has been applied during the simulation, as shown in Figure 13(b).The MVD dynamics is shown in Figure 13(d), it can be observed that for the values of air flow, and the variation of variables such as the number of conduits used, the velocity and temperature in the test section, as well as the LWC, the MVD values between and can be obtained.
Figures 14(a) and 14(b) show the flow of water through valves and , it can be seen how the PI control maintains the desired flow through them.Figures 14(c) and 14(d) show the opening in of the valves and through which the water flow passes.Furthermore, the behaviour of air through the valves and is shown in Figures 14(e) and 14(f), respectively.Whereas the opening of these valves is shown in Figures 14(g) and 14(h).The valve (water and air valves) are imposed an off and on behaviour at and , respectively.
The dynamics of the whole system is highly complex, involving a large number of variables; however, the developed model allows us to study our system under different scenarios and allows us to achieve values close to experimental data.
V Conclusions
In this paper, mathematical modelling based on first principles and machine learning techniques has been developed to form a hybrid model that allows us to perform an in-depth study of the atmospheric conditions in icing structures, using a subsonic wind tunnel as a facility.A statistical analysis of experimental data has been carried out in order to have a global idea of the behaviour of the system, which has allowed us to develop machine learning models of some variables of interest in the plant under study.The implemented model has been validated with real data and delivers valuable results.
The proposed methodology has allowed us to establish the dynamics of LWC and MVD in a simulation environment, as well as allowing us to make several variations in the parameters involved in the model, in order to study in depth the atmospheric conditions suitable for ice formation.
In future work, we intend to use optimisation techniques in order to optimise parameters in the system and improve the performance.In addition, we intend to develop intelligent control strategies and advanced control strategies such as fuzzy control, reinforcement learning, and model predictive control that allow us to control the desired LWC and MVD values minimising the use of system resources.
Acknowledgements
The authors would like to thank Vincent Sircoulomb, Associate Professor at ESIGELEC, for his help on the COPOGIRT project.
References
[1]S.Li and R.Paoli, “Aircraft icing severity evaluation,”Encyclopedia, vol.2, no.1, pp. 56–69, 2022.
[2]N.Davis, A.N. Hahmann, N.-E. Clausen, and M.Žagar, “Forecast of icingevents at a wind farm in sweden,” Journal of Applied Meteorology andClimatology, vol.53, no.2, pp. 262–281, 2014.
[3]S.Dierer, R.Oechslin, and R.Cattin, “Wind turbines in icing conditions:performance and prediction,” Advances in Science and Research,vol.6, no.1, pp. 245–250, 2011.
[4]F.Martini, L.T. ContrerasMontoya, and A.Ilinca, “Review of wind turbineicing modelling approaches,” Energies, vol.14, no.16, 2021.
[5]J.Molinder, H.Körnich, E.Olsson, and P.Hessling, “The use ofuncertainty quantification for the empirical modeling of wind turbineicing,” Journal of Applied Meteorology and Climatology, vol.58,no.9, 2019.
[6]R.W. Gent, N.P. Dart, and J.T. Cansdale, “Aircraft icing,”Philosophical Transactions of the Royal Society of London. Series A:Mathematical, Physical and Engineering Sciences, vol. 358, no. 1776, pp.2873–2911, 2000.
[7]Y.Cao, Z.Wu, Y.Su, and Z.Xu, “Aircraft flight characteristics in icingconditions,” Progress in Aerospace Sciences, vol.74, pp. 62–80,2015.
[8]R.E. Kreeger, L.Sankar, R.Narducci, and R.Kunz, “Progress in rotorcrafticing computational tool development,” SAE Technical Paper, 2015.
[9]H.Dobesch, D.Nikolov, and L.Makkonen, “Physical processes, modelling andmeasuring of icing effects in europe,” ÖsterreichischeBeiträge zu Meterologie und Geophysik, vol.34, 2005.
[10]J.Yang, W.Yu, J.Choisnard, A.Forcione, and S.Antic, “Coupledatmospheric–ice load model for evaluation of wind plant power loss,”Journal of Applied Meteorology and Climatology, vol.54, no.6, pp.1142–1161, 2015.
[11]I.Knop, S.E. Bansmer, V.Hahn, and C.Voigt, “Comparison of differentdroplet measurement techniques in the Braunschweig Icing Wind Tunnel,”Atmospheric Measurement Techniques, vol.14, no.2, pp. 1761–1781,2021.
[12]S.Rydblom and B.Thörnberg, “Measurement of atmospheric icing anddroplets,” IEEE Transactions on Instrumentation and Measurement,vol.69, no.8, pp. 5799–5809, 2020.
[13]C.R. Davison, J.D. MacLeod, and J.L. Chalmers, “Droplet evaporation modelfor determining liquid water content in engine icing tunnels and examinationof the factors affecting liquid water content,” in 9th AIAAAtmospheric and Space Environments Conference, 2017, p. 4246.
[14]MATLAB, version 9.13.0 (R2022b).Natick, Massachusetts: The MathWorks Inc., 2022.
[15]S.M. Krosel, G.L. Cole, W.M. Bruton, and J.R. Szuch, “A lumped parametermathematical model for simulation of subsonic wind tunnels,” NASA, Tech.Rep., 1986.
[16]B.Ilić, M.Miloš, M.Milosavljević, and J.Isaković,“Model-based stagnation pressure control in a supersonic wind tunnel,”FME Transactions, vol.44, no.1, pp. 1–9, 2016.
[17]J.D. Anderson, Modern Compressible Flow: With Historical Perspective,2nded., ser. Series in Aeronautical and Aerospace Engineering.New York: McGraw-Hill, 1990, vol.12.
[18]P.H. Oosthuizen and W.E. Carscallen, Introduction to Compressible FluidFlow, 2nded.Boca Raton, FL: CRCPress, 2013.
[19]Emerson, “Valve sizing calculation.”
[20]B.Ilić, M.Miloš, and J.Isaković, “Cascade nonlinearfeedforward-feedback control of stagnation pressure in a supersonic blowdownwind tunnel,” Measurement, vol.95, pp. 424–438, 2017.
[21]E.von Lavante, A.Zachcial, B.Nath, and H.Dietrich, “Unsteady effects incritical nozzles used for flow metering,” Measurement, vol.29,no.1, pp. 1–10, 2001.
[22]L.T. Vinh, S.Lee, Y.-T. Park, and B.J. d’Auriol, “A novel featureselection method based on normalized mutual information,” AppliedIntelligence, vol.37, pp. 100–120, 2012.
[23]C.Goutte and E.Gaussier, “A probabilistic interpretation of precision,recall and f-score, with implication for evaluation,” in Proceedingsof the European Conference on Information Retrieval.Springer, 2005, pp. 345–359.
[24]V.Rodriguez-Galiano, M.Sanchez-Castillo, M.Chica-Olmo, and M.Chica-Rivas,“Machine learning predictive models for mineral prospectivity: An evaluationof neural networks, random forest, regression trees and support vectormachines,” Ore Geology Reviews, vol.71, pp. 804–818, 2015.
[25]G.K. Tso and K.K. Yau, “Predicting electricity energy consumption: Acomparison of regression analysis, decision tree and neural networks,”Energy, vol.32, no.9, pp. 1761–1768, 2007.
[26]T.Chai and R.R. Draxler, “Root mean square error (rmse) or mean absoluteerror (mae)?–arguments against avoiding rmse in the literature,”Geoscientific model development, vol.7, no.3, pp. 1247–1250, 2014.
Introduction: My name is Tish Haag, I am a excited, delightful, curious, beautiful, agreeable, enchanting, fancy person who loves writing and wants to share my knowledge and understanding with you.
We notice you're using an ad blocker
Without advertising income, we can't keep making this site awesome for you.